# Quick-Start: Building Your Own Projects & Custom GPTs *Keystone Recognition · A companion to the AI Productivity Field Guide · prepared by Rogue Agents* You asked for a short, do-it-yourself guide to building the kind of reusable assistants we walked through together — the customer reply drafter, the company brain, the image tools. This is that guide. It's the same five-step process every time, on every platform. Keep it next to you the first couple of times you build one; after that you won't need it. --- ## First — which one are you building? Two things get confused because the names overlap. Pick based on what you want it *for*: - **A Project (Claude) or folder (ChatGPT)** — a *workspace* you keep for yourself. You load your background once and every chat inside it already knows your context. Best when you're organizing your *own* ongoing work. - **A Custom GPT (ChatGPT) or Gem (Gemini)** — a *tool you hand out*. Built to be shared by a link, even from a personal paid account. Best when other people need to use the same thing. Same engine underneath, same five steps to build. The only difference is whether it's a workspace you live in or a tool you give away. (If you're not sure, start with a Project for yourself — you can always rebuild it as a shareable Custom GPT later.) **You need a paid plan (~$20/month) on whichever platform you choose** — you can't build these on a free account, but you can build as many as you want once you're paid. --- ## The five steps (this is the whole thing) This is exactly what we did live when we built the reply drafter. Every assistant follows it. ### 1. Decide who it is and what it does Write one or two plain sentences, the way you'd brief a new hire on day one. Don't overthink it. > *"You're a customer service rep at Keystone Recognition. You draft friendly, professional email replies to customer questions. You never send anything — you hand a draft to a human to review."* ### 2. Give it your context This is what makes it *yours* instead of generic. Gather the background material it needs to know and load it as **knowledge files**. For Keystone that might be your About-Us / company brain, a customer list, a price list, product specs. - **Use markdown files (.md) when you can.** They're just clean text — light, fast, and easy for the tool to read. If you have a long PDF (a manual, a spec sheet), ask any AI to "convert this to a markdown file, keep everything the same" first, then upload that. - Spreadsheets and PDFs work too — but markdown is the tool's favorite. ### 3. Show it what good looks like This is the single biggest quality jump you can make. Give it real examples of the output you want. > For the reply drafter, we uploaded a file of example replies in Keystone's voice — a couple of white-glove ones, a couple of standard ones. A good example teaches more than a paragraph of instructions. If you don't have examples handy, ask the AI to *draft* a few for you, then edit them until they sound right, and use those. ### 4. Write the instructions Now put it together in the "instructions" box on the build screen. Keep it simple and delineated: - Who it is and what it does (from step 1) - The rules it must follow ("always confirm the order details back to the customer," "sign off with *Thanks, Keystone*," "never promise a delivery date you haven't confirmed") - How to use the knowledge files ("check the customer tier list to decide warm vs. standard tone," "use the company context file for background") > **Shortcut:** you can have the AI write these instructions *for* you. Tell it what you're building, answer its questions, and let it assemble the instructions — then paste them into the box. Using AI to help build the AI is fair game and faster. ### 5. Test it, then refine — then ship Save it and run it on a couple of **real** examples (for the reply drafter: one A-list customer email, one B-list). Check the output. It won't be perfect on the first try — that's normal. Adjust the instructions, add an example, tighten a rule, and run it again. Keep going until it's right about 95% of the time on the cases that matter. *Then* it's ready to use. --- ## Where the build screen lives The screen looks slightly different on each platform but holds the same things — a **name**, a short **description**, an **instructions** box, and a place to upload **knowledge files**. - **Claude (Project):** Projects → Create Project → add a name, set custom instructions, upload knowledge to the project's knowledge area. - **ChatGPT (Custom GPT):** your name (top right) → "My GPTs" → Create a GPT → use the *Configure* tab to set name, description, instructions, and upload knowledge. - **ChatGPT (Project/folder):** the Projects section in the sidebar — a workspace for your own chats, instructions, and files (the close cousin of a Claude Project). - **Gemini (Gem):** Gem manager → New Gem → name, instructions, and knowledge files; you can also set a default behavior (text, images, etc.). --- ## A worked example — your customer reply drafter This is the one we built live, written out so you can rebuild or extend it: 1. **Who it is:** *"You're a customer service rep at Keystone Recognition. You draft friendly, professional email replies. You never send — you hand a draft to a human to review."* 2. **Context:** upload (a) your company About-Us file, (b) a customer list sorted into tiers — A-list gets the warm, white-glove treatment, everyone else gets professional-but-confirming. 3. **Examples:** upload a handful of model replies — two written the white-glove way, two the standard way. 4. **Instructions:** *"Draft a reply in Keystone's voice. Look up the customer on the tier list. A-list → warm, reassuring, white-glove. Everyone else → professional and courteous, always confirm the key order details back to them. Sign off with* Thanks, Keystone. *Never promise a date you can't confirm."* 5. **Test:** paste in a real A-list email and a real B-list email, check both drafts, adjust, repeat. **Where to grow it next** (your idea in the session): add a library of canned explainers for the things you explain over and over — vector artwork requirements, why copy changes need an Excel and not a 70-page PDF — and let the drafter pull from those. Same five steps, just more examples loaded in. --- ## A few things to keep in mind - **Human in the loop, always.** These draft; they don't send. Read anything before it goes to a customer. - **Markdown for knowledge files.** Cleaner and lighter than PDF/HTML. Convert long documents first. - **Keep your source files in one place.** Store your knowledge files (the markdown ones) in a single tidy folder, organized by which tool they feed. When something changes, you update the file once and re-upload it. - **Sharing, for now:** Keystone doesn't have a shared team plan yet, so anything you build lives on your own account. To share an assistant across the team today, keep its knowledge files in one agreed shared folder and have each person load the same files into their own copy. A true shared team workspace is a small paid step you can take later, once a tool has proven it's worth it. - **It's a conversation, not a vending machine.** If a draft is off, tell it why and it'll do better. You can't break it. --- *Built one and want it to do more — or do it reliably at scale? That's the kind of thing Nate can take further. For now: pick the one tool you'd actually use this week, and build it.*